Dpd Import

/dpd-import

Announce: "Using dpd-import skill."

When to invoke

• User wants to analyze an external doc (spec, meeting notes, markdown, mermaid) against the current session
• User wants to run gap analysis on an existing document: /dpd-import <file> → /dpd-fill → /fcot
• User wants to import a prior decision set as an archived reference subgraph
• User says "この spec を取り込んで" / "外部 doc を DPD で分析したい"

Argument

/dpd-import <path-or-content> [--label=<label>]

<path-or-content>: file path or inline pasted content. --label: session/root label for the import (default: import-<filename>).

Supported input formats

Markdown (.md)

• Top-level headings (#) → root nodes or decision nodes
• Second-level headings (##) → sub-decision or plan nodes
• Decision verbs in body text ("decided", "will", "adopted", "rejected", "chose") → decision or resolution nodes
• Hypothesis-like text ("may", "might", "could", "option A/B/C") → hypothesis nodes
• References / citations / evidence markers → evidence nodes
• Constraint statements ("must", "shall not", "required") → constraint nodes
• Assumption statements ("assume", "we assume") → assumption nodes

Mermaid (.mmd or ```mermaid block)

• Parse node labels and edge types from the diagram
• Node shape hints: rectangles → plan/decision; diamonds → hypothesis; cylinders → evidence
• Edge labels: use as edge type where recognizable (e.g., -->|supports|)
• Unmapped shapes → question (open) as conservative default

YAML / JSON

• If the structure matches DPD node schema (type, text, children): import directly
• Otherwise: treat top-level keys as root topics, nested keys as child nodes, leaf values as text

Tool calls

1. Session setup

If no active session, or user wants a fresh import session:

start_session(scope=<sub-scope>, label="import-<doc-name>", mode='entry')

Otherwise, use the existing session.

2. Create import root

spawn_root(session_id=<session_id>, topic="Import: <doc-name>")

This root anchors all imported nodes. Record root_id.

3. Translate document to node/edge lists

Before calling any insert tool, translate the full document into:

nodes = [
  {parent_ref: <local ref>, type: <type>, text: <text>},
  ...
]
edges = [
  {from_ref: <local ref>, to_ref: <local ref>, type: <edge_type>},
  ...
]

Use local refs (e.g., sequential integers or heading slugs) to express structure before real node IDs exist. bulk_import_subgraph resolves them atomically.

4. Atomic import

bulk_import_subgraph(
  session_id=<session_id>,
  root_id=<root_id>,
  nodes=<nodes list>,
  edges=<edges list>,
  provenance='imported',
  state='archived'
)

All nodes receive provenance='imported' and state='archived'. This marks them as a hypothetical reference graph — not active work items.

Output

After import completes:

Import complete — <doc-name>
  Root: root_<id>
  Nodes imported: N
  Edges imported: M
  Provenance: imported / State: archived

Next steps:
  /dpd-fill   — detect gaps in the imported graph
  /fcot       — falsify imported decisions

/dpd-fill + /fcot orchestration

After import, suggest the full pipeline (§7.1):

/dpd-import <doc> → archived subgraph
/dpd-fill          → gap analysis (missing nodes, unstated assumptions)
/fcot              → falsify each archived decision

This pipeline enables systematic independent verification of an external spec. Suggest it explicitly after every import. User triggers each step manually.

For /fcot on imported decisions: invoke per decision node found in the imported graph. Call /fcot "<decision text>" for each high-stakes decision. Report results inline.

Eager edge-pinning to the focus subgraph (issue #41)

After bulk_import_subgraph returns, propose edges from the imported nodes to open questions / hypotheses in the focus subgraph before declaring the import done. Without this step, the imported graph sits in state='archived' and ends up being cited via free-text inside rationale content rather than typed edges — which undermines the value of having a graph at all for cross-source consistency (you can text-search but you cannot edge-walk).

Pattern:

1. Enumerate open question / hypothesis nodes in the focus subgraph (list_open_nodes, e.g. type="question" then type="hypothesis"). Since v0.11 this is paginated (50/page) — follow next_cursor until null to cover a large subtree.

2. For each, identify the top 1–3 imported nodes most likely to support, contradict, qualify, or provide derivation evidence (derived_from, supports, contradicts, qualifies).

3. Present as a bulk proposal:

   "Suggested cross-links (4 imported subgraphs → 3 focus open questions):

   • q_Q1 ← supports — n_imp_A4 (rate-limit guidance from §2.3)
   • q_Q2 ← derived_from — n_imp_B1 (canonical wire format)
   • q_Q3 ← qualifies — n_imp_C2 (clock-skew tolerance bound)

   Apply all / select / skip?"

4. Apply confirmed edges via add_edge.

Even when the enumeration is imperfect, the user can prune in one pass — that is cheaper than reconstructing edges during decision-formation, where missing edges silently degrade into prose citations.

Skip this step when the import is purely informational (e.g., glossary, change-log) with no clear attachment to current work. Default = propose.

Notes

• Imported nodes are state='archived' — they do not participate in active ambient signal detection. They serve as a reference layer.
• The supersedes edge mechanism can link imported decisions to active session decisions when a newer decision replaces an imported one: add_edge(from=<active_decision>, to=<imported_decision>, type='supersedes').
• Translation accuracy: document translation to node types is LLM inference. After import, run /dpd-fill to catch translation gaps before trusting the imported structure.
• bulk_import_subgraph is atomic: either all nodes + edges insert or none do. On error, report the raw error and ask user to correct the input.

Feedback footer

After completing a meaningful response (not for trivial status output), print exactly one line at the very end:

💬 Hit a bug or have feedback on DPD? Run /dpd-feedback "<short description>" or open an issue at https://github.com/o3co/agent-dpd/issues/new

This surfaces the dogfood feedback path without interrupting the main interaction. Keep it to one line. Do not repeat across multiple turns within the same exchange — once per skill invocation is enough.